def hyper_run_bm3d_tf(im_dir, im_name, sigma, nH, kH, NH, pH, tauMatchH,
useSDH, tau_2DH, lambda3DH, nW, kW, NW, pW, tauMatchW,
useSDW, lamb):
nW = 2 * kW
im = cv2.imread('test_data/image/' + im_name, cv2.IMREAD_GRAYSCALE)
im_noisy = cv2.imread(
'noisy_image_and_1st_res/' + im_name[:-4] + '_sigma' + str(sigma) +
'.png', cv2.IMREAD_GRAYSCALE)
im_basic = cv2.imread(
'noisy_image_and_1st_res/' + im_name[:-4] + '_sigma' + str(sigma) +
'_1st.png', cv2.IMREAD_GRAYSCALE)
im_noisy_p = symetrize(im_noisy, nW)
im_basic_p = symetrize(im_basic, nW)
im_denoised = bm3d_2nd_step_trendfilter3D(im_noisy_p, im_basic_p, nW, kW,
NW, pW, tauMatchW, useSDW, lamb)
im_denoised = im_denoised[nW:-nW, nW:-nW]
im_basic = (np.clip(im_basic, 0, 255)).astype(np.uint8)
im_denoised = (np.clip(im_denoised, 0, 255)).astype(np.uint8)
psnr = compute_psnr(im, im_denoised)
return im_basic, im_denoised, psnr
tauMatch_W = 400 if sigma < 35 else 3500 # ! threshold determinates similarity between patches
noisy_dir = '../test_data/sigma' + str(sigma)
im_path = os.path.join(input_dir, im_name)
im = cv2.imread(im_path, cv2.IMREAD_GRAYSCALE)
noisy_im_path = os.path.join(noisy_dir, im_name)
noisy_im = cv2.imread(noisy_im_path, cv2.IMREAD_GRAYSCALE)
# for lamb in [1, 2, 3, 4, 10, 20, 30, 40, 50, 60, 70, 80, 100]:
# for lamb in [7.2, 7.5, 7.7, 8, 8.2, 8.5, 8.7]:
# for lamb in [4, 5, 6, 7, 8, 8.3, 8.4, 9, 10, 12, 20]:
for lamb in [8.1, 8.2]:
im1, im2 = run_bm3d(noisy_im, sigma,
n_H, k_H, N_H, p_H, tauMatch_H, useSD_H, tau_2D_H, lambda3D_H,
n_W, k_W, N_W, p_W, tauMatch_W, useSD_W, tau_2D_W, (8.2, 8.2, lamb))
psnr_1st = compute_psnr(im, im1)
psnr_2nd = compute_psnr(im, im2)
im1 = (np.clip(im1, 0, 255)).astype(np.uint8)
im2 = (np.clip(im2, 0, 255)).astype(np.uint8)
save_name = im_name[:-4] + '+sigma_' + str(
sigma) + '-1st_' + tau_2D_H + '-lambda_' + str(lamb) + '=PSNR_' + '%.4f' % psnr_1st + '.png'
cv2.imwrite(os.path.join(save_dir, save_name), im1)
print(save_name)
save_name = im_name[:-4] + '+sigma_' + str(
sigma) + '-1st_' + tau_2D_H + '-2nd_' + tau_2D_W + '-lambda_' + str(
lamb) + '=PSNR_' + '%.4f' % psnr_2nd + '.png'
cv2.imwrite(os.path.join(save_dir, save_name), im2)
print(save_name)
if __name__ == '__main__':
from psnr import compute_psnr
from utils import add_gaussian_noise, symetrize
# <hyper parameter> -------------------------------------------------------------------------------
sigma = 20
nWien = 16
kWien = 8
NWien = 16
pWien = 3
tauMatchWien = 400 if sigma < 35 else 3500 # ! threshold determinates similarity between patches
useSD_w = True
tau_2D_wien = 'DCT'
# <\ hyper parameter> -----------------------------------------------------------------------------
img = cv2.imread('Cameraman256.png', cv2.IMREAD_GRAYSCALE)
img_noisy = cv2.imread('image_noise.png', cv2.IMREAD_GRAYSCALE)
img_basic = cv2.imread('y_basic.png', cv2.IMREAD_GRAYSCALE)
img_basic_p = symetrize(img_basic, nWien)
img_noisy_p = symetrize(img_noisy, nWien)
img_denoised = bm3d_2nd_step(sigma, img_noisy_p, img_basic_p, nWien, kWien,
NWien, pWien, tauMatchWien, useSD_w,
tau_2D_wien)
img_denoised = img_denoised[nWien:-nWien, nWien:-nWien]
psnr_2st = compute_psnr(img, img_denoised)
print('img and img_denoised PSNR: ', psnr_2st)
# cv2.imwrite('y_final.png', img_denoised.astype(np.uint8))
def compare_psnr(im_name):
sigma_list = [2, 5, 10, 20, 30, 40, 60, 80, 100]
im_dir = '../test_data/image'
py_res_dir = 'python'
cpp_res_dir = 'cpp'
im = cv2.imread(os.path.join(im_dir, im_name))
f_p_c_1 = list()
u_p_c_1 = list()
f_p_c_2 = list()
u_p_c_2 = list()
f_p_p_1 = list()
u_p_p_1 = list()
f_p_p_2 = list()
u_p_p_2 = list()
for sigma in sigma_list:
for cpp_res_im_name in os.listdir(cpp_res_dir):
if im_name[:-4] + '_s' + str(sigma) + '_' in cpp_res_im_name:
if 'cpp_1st' in cpp_res_im_name:
P_str = cpp_res_im_name.split('_')[-1][1:-4]
float_psnr_cpp_1st = float(P_str)
f_p_c_1.append(float_psnr_cpp_1st)
im_res = cv2.imread(
os.path.join(cpp_res_dir, cpp_res_im_name))
uint8_psnr_cpp_1st = compute_psnr(im, im_res)
u_p_c_1.append(uint8_psnr_cpp_1st)
elif 'cpp_2nd' in cpp_res_im_name:
P_str = cpp_res_im_name.split('_')[-1][1:-4]
float_psnr_cpp_2nd = float(P_str)
f_p_c_2.append(float_psnr_cpp_2nd)
im_res = cv2.imread(
os.path.join(cpp_res_dir, cpp_res_im_name))
uint8_psnr_cpp_2nd = compute_psnr(im, im_res)
u_p_c_2.append(uint8_psnr_cpp_2nd)
for py_res_im_name in os.listdir(py_res_dir):
if im_name[:-4] + '_s' + str(sigma) + '_' in py_res_im_name:
if 'py_1st' in py_res_im_name:
P_str = py_res_im_name.split('_')[-1][1:-4]
float_psnr_py_1st = float(P_str)
f_p_p_1.append(float_psnr_py_1st)
im_res = cv2.imread(
os.path.join(py_res_dir, py_res_im_name))
uint8_psnr_py_1st = compute_psnr(im, im_res)
u_p_p_1.append(uint8_psnr_py_1st)
elif 'py_2nd' in py_res_im_name:
P_str = py_res_im_name.split('_')[-1][1:-4]
float_psnr_py_2nd = float(P_str)
f_p_p_2.append(float_psnr_py_2nd)
im_res = cv2.imread(
os.path.join(py_res_dir, py_res_im_name))
uint8_psnr_py_2nd = compute_psnr(im, im_res)
u_p_p_2.append(uint8_psnr_py_2nd)
l1 = plt.plot(sigma_list, u_p_c_2, 'g--', label='cpp_version')
l2 = plt.plot(sigma_list, u_p_p_2, 'b--', label='py_version')
for x, cpp, py in zip(sigma_list, u_p_c_2, u_p_p_2):
plt.text(x,
cpp,
'cpp' + '%.3f' % cpp,
ha='center',
va='top',
fontsize=10,
color='g')
plt.text(x,
py,
'py' + '%.3f' % py,
ha='center',
va='bottom',
fontsize=10,
color='b')
plt.rcParams['savefig.dpi'] = 300 # 图片像素
plt.rcParams['figure.dpi'] = 300 # 分辨率
plt.title(im_name)
plt.xlabel('sigma')
plt.ylabel('psnr')
plt.legend()
plt.savefig(im_name)
plt.show()
